Intelligent Identification Method for Vibration Characteristics of Rotating Machinery

Abstract

The present invention relates to an intelligent identification method for vibration characteristics of rotating machinery, which comprises the following steps in order: Step 1, converting the velocity or acceleration time domain signal of the mechanical vibration into a frequency domain envelope spectrum by signal processing, and extracting the frequency upper limit value f of the envelope spectrum. max ; Step 2, filter out at least the frequency range of f by amplitude comparison max / N max High energy harmonics within, where N max The upper limit multiple of the frequency multiplication for the frequency multiplication verification of high energy harmonics; Step 3, extract 1 to N of each high energy harmonic in turn max At least one set of characteristic parameters based on the amplitude and / or frequency of the peaks in the frequency-doubled region, where the peak of the high-energy harmonic in the frequency-doubled region is the high-energy harmonic itself; and step 4, inputting the at least one set of characteristic parameters of each high-energy harmonic into a machine learning intelligent algorithm for training and calculation. The above intelligent identification method can not only significantly improve the efficiency and accuracy of mechanical fault identification, but also enable real-time online monitoring of rotating machinery.

Description

Technical Field

[0001] The present invention relates to an intelligent recognition method for vibration characteristics of rotating machinery based on envelope spectrum, and in particular to an intelligent diagnosis method for faults of rotating machinery. Background Art

[0002] Frequency domain analysis is a common method for diagnosing mechanical faults, which is generally achieved by converting mechanical vibration signals (such as velocity or acceleration time domain signals) into frequency domain signals. Mechanical faults in rotating machines, such as damage to the rotor, gears or bearings, will cause systematic oscillations or shocks. The frequency response in the envelope spectrum reflects the nature and intensity of such oscillations and shocks, and can therefore be used to accurately identify various faults in rotating machines. Specifically, vibrations and shocks induce a defect harmonic family in the spectrum. The harmonic family includes fundamental defect harmonics whose frequencies correspond to the characteristic frequencies of the mechanical faults and high-frequency defect harmonics whose frequencies are integer multiples of the fundamental defect harmonic frequencies. The frequencies and amplitudes of the fundamental defect harmonics and high-frequency defect harmonics are directly related to the type and severity of the mechanical fault, making them the most commonly used and intuitive identification features in mechanical fault diagnosis.

[0003] Currently, fault harmonic diagnosis relies primarily on manual labor. Limited by individual skills and experience, its reliability and accuracy are insufficient to meet current needs. Furthermore, manual diagnosis cannot monitor equipment in real time. Faults in their early stages are often difficult to detect with the naked eye, making it easy to miss optimal maintenance opportunities and leading to accidents. Reality calls for an automated diagnostic technology that is both accurate and reliable, and capable of real-time online monitoring. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides an intelligent identification method for the vibration characteristics of rotating machinery, which comprises the following steps in sequence: Step 1, converting the velocity or acceleration time domain signal of the rotating machinery vibration into a frequency domain envelope spectrum through signal processing, and extracting the frequency upper limit value f of the envelope spectrum. max ; Step 2, filter out at least the frequency range of f by amplitude comparison max / N max High energy harmonics within, where N max The upper limit multiple of the frequency multiplication for the frequency multiplication verification of high energy harmonics; Step 3, extract each high energy harmonic 1 to N in turn maxAt least one set of characteristic parameters of the peak values ​​in the frequency doubling region based on their respective amplitudes and / or frequencies, wherein the peak value in the frequency doubling region of the high-energy harmonic is the high-energy harmonic itself; and, step 4, inputting the at least one set of characteristic parameters of each high-energy harmonic into a machine learning intelligent algorithm for training and calculation one by one.

[0005] The use of this intelligent identification method not only significantly improves the efficiency and accuracy of mechanical fault identification but also enables real-time online monitoring of rotating machinery. This intelligent detection method ensures that rotating machinery is always under safe monitoring during operation, which has great practical significance and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 It is a time domain signal diagram of the speed of a rotating mechanical vibration;

[0007] Figure 2 for Figure 1 The frequency domain envelope spectrum of the velocity time domain signal is shown;

[0008] Figure 3 is the frequency domain envelope spectrum in units of order;

[0009] Figure 4 This is a schematic diagram of the bamboo grass filtering method;

[0010] Figure 5 Schematic diagram of a method for verifying the harmonic relationship of the double frequency of high-energy harmonics;

[0011] Figure 6 It is a model diagram of a fully connected neural network;

[0012] Figure 7 is a graph showing the relationship between variables of the sigma adjustment function used in the present invention; and

[0013] Figure 8 This is a comparison chart of the relationship between the envelope spectrum and the corresponding bamboo grass proportional coefficient. DETAILED DESCRIPTION

[0014] Since the velocity and acceleration signals contain most of the key information of the mechanical vibration, the present invention chooses to convert the velocity or acceleration time domain signal of the mechanical vibration into a frequency domain envelope spectrum through demodulation. Figure 1 Displays the time domain signal of the vibration velocity of a rotating machine in a stable state, where the horizontal axis represents time (s) and the vertical axis represents acceleration (m / s). After Fourier transform, the time domain signal is converted into Figure 2 The frequency domain envelope spectrum shown in FIG, wherein the horizontal axis represents the frequency (Hz) and the vertical axis represents the envelope value (vE), wherein vE is the virtual value obtained after the envelope transformation of the velocity unit m / s. Figure 1If it is an acceleration time domain signal, the frequency unit of the horizontal axis in its corresponding envelope spectrum is Hertz (Hz), and the vertical axis unit is acceleration (m / s 2 ) is the virtual unit (gE) obtained after envelope transformation (not shown in the figure).

[0015] The envelope spectrum also has a frequency unit that is a multiple of the machine speed, called "order", such as Figure 3 As shown. The order can reflect the relationship between the frequency multiples (referred to as "multiples") between the fault harmonics and the machine speed, which is easy to compare and identify with the fault characteristic frequency, thus facilitating the judgment of the fault nature. Figure 3 In the figure, some frequencies have signal amplitudes significantly higher than others. These are defined as "high-energy harmonics" in this disclosure and are marked with an asterisk (*). High-energy harmonics arise from significant vibrations generated by the machine during operation and may be members of a fault harmonic family. In this disclosure, they are identified and tested as suspected fault harmonics.

[0016] The identification of high-energy harmonics in the present invention is achieved through amplitude comparison, that is, by comparing the amplitude of a spectral line sample with the amplitude of the adjacent spectral line sample to screen out signal individuals with sufficiently high energy. Screening for high-energy harmonics can be achieved by looking for local peaks within the spectrum range. If the height (envelope value) of a spectral line is greater than the envelope value of the adjacent spectral line, then it is defined as a "local peak" in the present invention. The "adjacent" can be defined by the number of spectral lines before and after a specific spectral line sample in the present invention. For example, the range defined by the front and back k spectral lines is "adjacent", where k is a natural number, then when k=1, as long as a sample signal is larger than the amplitudes corresponding to the two spectral lines on its left and right, then the sample signal is identified as a local peak. In summary, a local peak is a sample individual with the largest amplitude within the range defined by the front and back k spectral lines.

[0017] While it's feasible to use local peaks as suspected fault harmonics for subsequent testing, this approach is computationally inefficient. This is because, while local peaks are "peaks," they aren't necessarily "high energy." Fault harmonics often have significantly higher energy amplitudes than noise signals. Therefore, high-energy harmonics significantly higher than those of adjacent signals are most likely to be fault harmonics. Therefore, it's necessary to further filter out high-energy harmonics with significantly higher amplitudes than adjacent signals from the local peaks; these are the most likely fault harmonics to be found and identified.

[0018] The process of further filtering out high-energy harmonics (suspected fault harmonics) from local peaks is referred to as "bamboo filtering" in this invention. This is a metaphor for finding bamboos (high-energy harmonics) with amplitudes significantly higher than the average height of the grass (noise). Figure 4A specific implementation method of bamboo grass filtering is to set a window within the range of j spectral lines before and after the local peak (j is a natural number), and calculate the average amplitude mean (A) of all spectral line samples in the window except the local peak. -j :A j ), and then calculate the amplitude of the local peak A0 and the average amplitude of the adjacent samples mean(A -j :A j ) to obtain the ratio of bamboo grass ratio BambooGrassRatio=A0 / mean(A -j :A j ). Among them, a preferred implementation is to set j ≥ k, so that the comparison range of the bamboo grass ratio is larger than the comparison range of the local peak, which helps to improve the recognition accuracy of high-energy harmonics. It is not difficult to understand that the bamboo grass ratio generally reflects the degree to which the local peak protrudes from the adjacent samples. Only local peaks with amplitudes significantly higher than those of the adjacent samples are considered high-energy harmonics. To facilitate screening, a threshold value can be set for the bamboo grass ratio, such as 3.5. Only local peaks above this threshold value are considered high-energy harmonics, otherwise they are considered noise.

[0019] exist Figure 4 In the embodiment shown, the bamboo grass ratio is based on the average value of the neighboring samples of the local peak as the comparison basis (denominator), that is, the proportion of each neighboring sample in the selected window in the average value is the same. This is inconsistent with the actual situation that the amplitudes of the neighboring samples of the fault harmonic decay in sequence, which can easily lead to misjudgment and omission of high-energy harmonics. For example, if two high-energy harmonics are just adjacent to each other and are within the selected window (j spectral lines) of each other, then this will inevitably cause the average amplitude mean (A) of the neighboring samples within the selected window to be equal to 0. -j :A j ) increases, resulting in a decrease in the proportion of bamboo grass corresponding to high-energy harmonics. When the bamboo grass proportion is lower than the set threshold, the corresponding high-energy harmonics will be missed.

[0020] To avoid the above situation, a feasible implementation method is to redistribute the weights of the neighboring samples of high-energy harmonics in the average value algorithm. For example, a triangular distribution or normal distribution model (not shown in the figure) can be used in the selected window to give higher weights to the neighboring samples close to the local peak and lower weights to the neighboring samples far from the local peak, so that the average amplitude mean (A) of the neighboring samples used as the comparison basis is -j :A j ) is expanded from the mathematical mean to the weighted mean.

[0021] In the above specific implementation, the mathematical mean or weighted mean (A -j :Aj ) does not comprise local peak A0 itself. Yet those skilled in the art will appreciate that it is also feasible to include local peak A0 in the calculation process of mean value. At this moment, as long as the empirical threshold value of the bamboo grass ratio is adjusted, high-energy harmonics can still be screened out on the basis of said mean value.

[0022] The process of screening high-energy harmonics within the spectrum by amplitude comparison has been described above. Among them, the combined use of local peak and bamboo grass filtering is an effective method for realizing the rapid screening of high-energy harmonics. However, it should be understood by those skilled in the art that any method based on amplitude comparison can achieve the purpose of high-energy harmonic screening according to the present invention, and is not limited to the use of the above-mentioned specific method. For example, processing the data of all spectral line samples by bamboo grass green wave alone can also realize the accurate identification of high-energy harmonics, but the computational efficiency is not as fast and convenient as carrying out bamboo grass filtering on the basis of local peak.

[0023] As is well known, the envelope spectrum is a discrete spectrum obtained by Fourier transforming a time domain signal. Its spectral line samples often deviate from the real signal in amplitude, frequency and phase. For example, due to the limitation of the sampling frequency, the real local peak should actually be between the local peak and the local sub-peak, and the amplitude should also be higher than the local peak. There are many methods for correcting the discrete spectrum in existing theories, such as the ratio correction method, the energy center of gravity correction method, etc., which will not be expanded here. As a preferred embodiment, the present invention can choose to make corrections to the local peaks in the spectral line samples and replace the original spectral line sample data with the corrected data, so that the data accuracy of the local peak can be greatly improved without increasing the sampling frequency (spectral line density), which is very beneficial to improving the accuracy of the subsequent machine learning intelligent algorithm for fault harmonic judgment. Considering that all high-energy harmonics are screened from the local peak, the correction process should be arranged after the screening process of the local peak and before the bamboo grass filtering process, so as to ensure that the screened high-energy harmonics automatically have accurate amplitude, frequency and phase.

[0024] The following describes the process of using the learning ability of the intelligent algorithm to determine whether the high-energy harmonic in the envelope spectrum is a member of the fault harmonic family. Figure 3As shown, the typical distribution of a fault harmonic family in the spectrum is a series of fault harmonic components whose frequencies are integer multiples. Furthermore, as the frequency multiple increases, the amplitude of the fault harmonic components decreases. Data with the aforementioned characteristics is used as positive samples, and data without these characteristics is used as negative samples. The large number of collected positive and negative samples is randomly divided into training and test sets according to a certain ratio. The training set is then used to train the machine learning intelligent algorithm model, and the test set is used to test the trained model. If the test results are sufficiently good, it can be put into practical use.

[0025] Figure 5 The following is a schematic diagram of a method for verifying the harmonic relationship between high-energy harmonics and their double frequency. Assume that the frequency corresponding to the first high-energy harmonic on the left side of the figure is f1, and its amplitude is y1. If this high-energy harmonic happens to be the first harmonic of a fault harmonic family (i.e., the "fundamental harmonic"), then the envelope spectrum should contain a second harmonic from the same harmonic family (i.e., the "double frequency harmonic") near its double frequency 2f1 region. Taking into account the existence of errors, the verification method is to search for the double frequency region peak y2 = max(2f1) within a certain (window) range before and after the double frequency 2f1, and then extract the amplitude y2 and frequency f2 corresponding to this double frequency region peak (f2 ≈ 2f1).

[0026] Taking the double frequency region window defined by m spectral lines before and after as an example (m is a natural number), the double frequency region peak y2 should be the maximum value max (ENV -m :ENV m ), where m is preferably greater than or equal to j, and ENV represents the envelope value of the spectral line sample. If the double frequency region peak (f2, y2) found according to the above method corresponds to a high-energy harmonic screened out before, then the two high-energy harmonics corresponding to the frequencies f1 and f2 are most likely the first and second harmonics in the same fault harmonic family. On the contrary, if the double frequency region peak (f2, y2) found according to the above method does not correspond to any high-energy harmonic screened out before, then the current high-energy harmonic (f1, y1) and its double frequency region peak (f2, y2) are most likely not the first and second harmonics in the same fault harmonic family. Of course, the above judgment process is analyzed and described according to the logic of human thinking. In the present invention, the above judgment process is completed by machine calculation in subsequent steps by a machine learning intelligent algorithm. Specifically, a set of frequencies (f1, f2) and / or a set of regional peaks (y1, y2) corresponding to the peaks in the first-harmonic region (i.e., the high-energy harmonic itself) and the second-harmonic region are extracted as input items of the intelligent method and obtained by the machine execution algorithm.

[0027] It is not difficult to understand that for a specific high-energy harmonic, it may not be sufficient to only verify whether the peak value in the first harmonic region (i.e., the high-energy harmonic itself) and the peak value in the second harmonic region have a unique relationship with the fault harmonic component. For the sake of accuracy, it is necessary to at least perform the above verification on the peak value in the third harmonic region, that is, to search for the peak value in the third harmonic region (f3, y3) within a certain frequency range before and after the third harmonic frequency 3f1, where f3 ≈ 3f1. Then, a set of frequency data (f1, f2, f3) and / or a set of regional peak data (y1, y2, y3) corresponding to the peak values ​​in the first to third harmonic regions of the high-energy harmonic are extracted, and the data are used as input items for subsequent machine learning intelligent methods, or as the parameter basis for further data processing.

[0028] However, the more verification of the fault harmonic relationship, the better. Generally speaking, the harmonic components above ten times the frequency may have been completely attenuated and almost completely submerged in the noise signal (grass) and difficult to identify. Not only that, the verification of the high-frequency relationship will also be over-amplified due to the error of the basic harmonic frequency f1, making the above method of finding the regional peak according to the frequency multiplication relationship invalid due to the large frequency multiplication error. Therefore, the frequency multiplication verification of high-energy harmonics should not exceed the ten times the frequency region, and at most the frequencies corresponding to the peaks of the one to ten times the frequency region (f1, f2...f 10 ) and / or amplitude (y1, y2...y 10 ) as the basis for subsequent judgment or data processing. However, considering the limitations of computer processing power and frequency multiplication errors, it is considered more appropriate to test the fault harmonic relationship of high-energy harmonics with an upper limit of 4-7 times the frequency, and the test of the fault harmonic relationship with an upper limit of 5-6 times the frequency has been proven to be the most effective.

[0029] The above targets Figure 5 The process of fault harmonic inspection for the first high-energy harmonic on the left is completely applicable to the second high-energy harmonic on the left. Similarly, the high-energy harmonics in the spectrum can be verified one by one according to the above method until the upper limit frequency N of the frequency multiple inspection of the next high-energy harmonic is reached. max f1 exceeds the upper limit of the frequency range f of the spectrum max So far, where N max It is obvious that the method of the present invention for testing the fault harmonic relationship of high energy harmonics must use the upper limit frequency N of the high energy harmonic frequency multiplication test. max f1 falls at the upper frequency f of the spectrum max In other words, the upper limit frequency f of the spectrum must be max The upper limit of the frequency multiplication of high energy harmonic frequency multiplication test N max The ratio f max / N maxThe condition is that it can at least cover the fundamental harmonic frequency f1 of common mechanical failures, that is, satisfy f1≤f max / N max or f max ≥N max f1, otherwise the upper frequency f of the spectrum will have to be expanded max Or choose a smaller upper limit multiple N of the frequency test max To make the upper limit frequency N of the high energy harmonic frequency multiplication test max f1 is always at the upper frequency limit f of the spectrum max Within the range.

[0030] Figure 6 This is a typical model diagram of a fully connected neural network, a common type of machine learning intelligent algorithm mentioned above. A fully connected neural network generally adopts a pyramidal structure, with the input layer at the bottom on the left and the output layer at the top on the right. Between the bottom and top are several hidden layers. The input layer contains a number of input units corresponding to the number of input data items, and the output layer contains at least one output unit. For example, if the output value is 1, the diagnosis is YES, indicating that a faulty harmonic relationship has been identified; if the output value is 0, the diagnosis is NO, indicating that no faulty harmonic relationship has been identified. The number of hidden layers and the number of units in each layer should be consistent with the number of input data items.

[0031] Taking the inspection of the fault harmonic relationship of the 5-fold frequency upper limit of high-energy harmonics as an example, according to the method described above, a set of frequency data (f1, f2...f5) and / or a set of amplitude data (y1, y2...y5) corresponding to the peak values ​​of the 1 to 5-fold frequency region of high-energy harmonics must be extracted. One implementation method is to input a set of frequency data (f1, f2...f5) into a neural network with an input layer having five input units. The trained neural network can then give an output value at the output unit to determine whether this set (of 5 items) of frequency data corresponds to a set of fault harmonic families. Another implementation method is to input a set of amplitude data (y1, y2...y5) into a neural network with an input layer having five input units. The trained neural network can then give an output value at the output unit to determine whether this set (of 5 items) of amplitude data corresponds to a set of fault harmonic families. In order to improve the accuracy of intelligent diagnosis, a preferred implementation method is to simultaneously input the above-mentioned set of (5 items) frequency and set of (5 items) amplitude (a total of 10 items) data into a neural network with 10 input units. The trained neural network can then give an output value at the output unit to determine whether the peak value of the 1-5th frequency region of the high-energy harmonic corresponding to these two sets of (a total of 10 items) data has a fault harmonic family relationship unique to mechanical faults.max After all high-energy harmonics within the / 5 frequency range are tested one by one, as long as the diagnosis result of a group of high-energy harmonics is YES, it indicates the existence of a mechanical fault harmonic group, and the system is diagnosed as having a fault.

[0032] Given the powerful functions of neural networks, the data used as input items are not limited to the above-mentioned frequency and / or amplitude itself, but can also be derived data of the two, including but not limited to their respective function values, statistical values, comparative values, etc. Taking frequency as an example, its derived data can be the frequency deviation dev(f n )=|f n -nf1| / f width , where f width is the bandwidth of the spectral line interval, n is 1 to N max Integer between . The frequency deviation dev(f n ) refers to the actual frequency f corresponding to the peak value in the frequency doubling region. n The frequency deviation from the expected fault harmonic frequency nf1 relative to the bandwidth f width The ratio of the frequency deviation dev(f n ) value is larger, indicating that the peak frequency f in the frequency doubling region n The greater the deviation between the frequency multiple nf1 and the expected high-energy harmonic, the less likely the high-energy harmonic being tested is a fault harmonic; the frequency multiple deviation dev(f n ) value is smaller, indicating that the peak frequency f in the frequency doubling region n The smaller the deviation between the frequency multiple nf1 and the expected high-energy harmonic, the greater the possibility that the high-energy harmonic being tested is the fault harmonic. As a derivative of the frequency, a set of frequency multiple deviations The data can be used independently or additionally as input items of the neural network to reflect whether the peak value of the frequency region and the measured high-energy harmonic have a frequency multiple relationship unique to the members of the fault harmonic family from the dimension of frequency multiple matching degree.

[0033] It is not difficult to see that the frequency deviation dev(f n ) is a relative value function, and its size is related to the denominator used as a comparison benchmark. Although the above dev(f n ) in the expression of the spectral line width f width It is used as the basis for comparison (denominator), but choosing other scales, such as setting the width of the window, as the basis for comparison is also feasible and does not affect the calculation results of the machine learning algorithm at all.

[0034] It is also necessary to point out that when n = 1, dev(f1) = 0. This is because the high-energy harmonic and its first harmonic peak are actually the same signal, and there is no frequency deviation between the two. Therefore, using the first harmonic deviation dev(f1) with a value of zero as the input of the intelligent algorithm is actually ineffective in determining the presence of the fault harmonic family. Therefore, the set of harmonic deviation data input into the neural network can contain only dev(f1) instead of dev(f1). This can reduce the occupation of one input unit of the neural network. However, considering the integrity of the data, a complete set of octave deviation data It is also feasible to use it as an input item for intelligent algorithms and will not have a substantial impact on the judgment results.

[0035] From the above frequency deviation dev(f n ) formula, it can be seen that the method of testing high-energy harmonics by frequency and / or its derivative data (hereinafter collectively referred to as "frequency-based characteristic parameters") will increase the deviation as the number of multiplications n increases. This is because the frequency f of the peak in the multiplication region is n The deviation between the expected fault harmonic frequency nf1 and the multiplication frequency nf1 will gradually increase with the increase of the multiplication number, which is not conducive to the machine learning intelligent method to make a correct judgment. This explains why the invention sets the upper limit multiple N of the multiplication test for high-energy harmonics from the perspective of multiplication deviation. max The reason it's not advisable to set the value too high also explains the necessity of correcting all local peaks within the spectrum to minimize errors. In this case, all local peaks are corrected, and the high-energy harmonics detected are also corrected. Corrected high-energy harmonics have characteristic parameters that are closer to reality, greatly facilitating accurate judgment by machine learning intelligent methods during frequency multiplication testing.

[0036] Similar to the case of frequency, amplitude can also have its own derivative data. For example, the ratio of bamboo to grass mentioned above is a typical amplitude derivative data. max Peak value in the octave region The ratio of the data to the (weighted) mean of its neighboring samples This forms a set of amplitude derivative data. Such a set of amplitude derivative data can be used independently or additionally as input items of the neural network to reflect the 1 to N high-energy harmonics from the dimension of relative height. max The degree to which the peak in the octave region stands out from the neighboring samples in a specified window. Similar to the case of frequency, amplitude and / or its derivative data are referred to as "amplitude-based feature parameters" in the present invention.

[0037] It is not difficult to understand that the bamboo-grass ratio is not the only form of derived data of amplitude. In the field of artificial intelligence, the sigma function (Sigmoid(x)=1 / (1+e -x )=e x / (e x +1)) is often used to convert a large-scale numerical value x into a smaller-scale function value Sigmoid(x). As a specific embodiment, the present invention actually uses the bamboo grass ratio adjustment function S(BambooGrassRatio)=Sigmoid(2.4·BambooGrassRatio-6.6) to further "compress" the bamboo grass ratio BambooGrassRatio, which has a numerical range of zero to several tens, into a bamboo grass ratio Sigma adjustment function S with an amplitude range of 0-1. The latter is actually a coefficient formed by mathematically adjusting the bamboo grass ratio, and is therefore also referred to as the "bamboo grass ratio coefficient" in the present invention.

[0038] Figure 7 This is a variable relationship diagram of the bamboo grass ratio sigma adjustment function actually adopted in the present invention, wherein the horizontal axis represents the bamboo grass ratio BambooGrassRatio and the vertical axis represents the bamboo grass ratio coefficient S. As can be seen from the figure, when the value of the bamboo grass ratio coefficient S approaches 1, it indicates that the regional peak is significantly higher than the (weighted) average value of the adjacent samples in the set window; when the value of the bamboo grass ratio coefficient S approaches 0, it indicates that the regional peak is not significantly higher than the (weighted) average value of the adjacent samples in the set window; the intermediate value (threshold) between significant and insignificant is set to 2.75, that is, when the bamboo grass ratio BambooGrassRatio is equal to 2.75, the bamboo grass ratio coefficient S just takes the intermediate value 0.5.

[0039] Figure 8 The following is a comparison chart of the relationship between the envelope spectrum and the corresponding bamboo grass ratio coefficient. As can be seen from the figure, the bamboo grass ratio coefficients marked with dotted lines are mostly distributed near the "polar" areas with values ​​close to 0 and 1. The above characteristics actually come from the polarization "stretching" effect of the sigma function. Therefore, compared with the bamboo grass ratio, the bamboo grass ratio coefficient as an input item of the machine learning intelligent algorithm is more conducive to the classification judgment of the machine intelligence algorithm. Therefore, a set of bamboo grass ratio coefficients Independently or additionally as an input item of the neural network, it can also reflect the degree to which the peak of the high-energy harmonic frequency region protrudes from the adjacent samples in the set window from the dimension of relative height.

[0040] As a preferred embodiment, the present invention can simultaneously convert 1 to N high energy harmonics max A set of amplitude data corresponding to the peak value in the frequency doubling region A set of bamboo grass ratio coefficient data and a set of octave deviation data As the input of the neural network, it judges whether the current high-energy harmonic is a fault harmonic from the three dimensions of absolute amplitude, relative amplitude and frequency matching, thereby greatly improving the accuracy of intelligent identification.

[0041] The above describes the entire process of mechanical fault diagnosis using a neural network as an example of an intelligent algorithm, which uses the characteristic parameters of high-energy harmonics as input into a machine learning intelligent algorithm. It is readily understood that any other type of intelligent algorithm, such as logistic regression or random forest, can achieve the intelligent identification objectives described in this invention as long as it can learn from training data samples.

[0042] Generally speaking, the vast majority of faults in rotating machinery are related to their rotors, such as rotor imbalance and misalignment. Therefore, the rotor's rotational frequency can be used to verify the accuracy of intelligent identification results. Assuming the intelligent identification result is YES, indicating that a certain high-energy harmonic has a corresponding fault harmonic family, the correctness of the intelligent identification result can be verified by comparing the high-energy harmonic with the rotor's rotational frequency. For example, by comparing the frequency of the high-energy harmonic with the bearing's fundamental fault frequencies, such as the ball pass frequency inner race (BPFI), the ball pass frequency outer race (BPFO), and the rolling element fault frequency (BPF), it is possible to further confirm whether the fault originates from the bearing. Similarly, by comparing the frequency of the high-energy harmonic with gear fault fundamental frequencies, such as the gear rotation frequency, gear meshing frequency, and gear natural frequency, it is possible to determine whether the fault originates from the gear system. It can be seen that the characteristic frequency of the rotating machine can not only be used to verify the results of intelligent identification, but also help to determine the nature and source of the fault, so as to facilitate targeted preventive or maintenance measures.

[0043] As can be seen from the above description, using fault characteristic frequencies to verify intelligent identification results is not limited to machine components such as rotors, bearings, and gears; in theory, it should be applicable to all types of rotating machinery. In the above verification process, the frequency f1 of the high-energy harmonic corresponding to the equipment's fundamental fault frequency is the characteristic parameter required for this verification. Of course, this does not prevent a set of frequency data (f1, f2, ..., f5) from being used as input for the machine learning intelligent algorithm for the previously described high-energy harmonic frequency multiplication test.

[0044] Mechanical vibrations are bound to manifest as characteristic parameters in the spectrum, regardless of whether a mechanical fault is included therein. In this sense, the method of identifying mechanical faults based on characteristic parameters in the spectrum of the present invention can actually have a wide range of uses, and is not limited to the identification and diagnosis of mechanical faults. Taking vibration and noise as an example, vibration causes noise, but noise is not necessarily caused by faults. The present invention utilizes the learning ability of intelligent algorithms to train the algorithm model through a training set, and can automatically match and associate human auditory perception of noise with characteristic parameters in the spectrum, thereby guiding people to have a new understanding of the source and generation mechanism of noise. In this sense, the present invention is actually an intelligent method for judging and identifying the causes and mechanisms of the vibration characteristics based on the vibration characteristics of a rotating machine.

[0045] Taking the bearing field as an example, noise is a common problem. The present invention can be used to identify bearing noise or analyze the mechanism by which bearings generate noise. Current bearing noise standards focus only on a few indicators, such as the bearing's vibration velocity peak and vibration velocity crest factor. However, these indicators may not be the sole source of bearing noise. The present invention utilizes intelligent algorithms based on characteristic parameters in the frequency spectrum to transcend current human cognition, leading people to explore the underlying mechanisms of noise generation, thereby enabling targeted measures to be taken in practice to reduce bearing noise levels.

[0046] Those skilled in the art will appreciate that the intelligent identification method for characteristic parameters is not limited by the specific implementation. Any changes and improvements to the present invention, as long as they comply with the limitations of the appended claims, fall within the scope of protection of the present invention.

Claims

1. A method for intelligently identifying vibration characteristics of a rotating machine, comprising the following steps in the following order: Step 1: Convert the velocity or acceleration time domain signal of the mechanical vibration into the frequency domain envelope spectrum through signal processing, and record the frequency upper limit value f of the envelope spectrum. max ; Step 2: By comparing the amplitude, at least the frequency range of f is selected. max / N max High energy harmonics within, where N max The upper limit multiple of the frequency multiplication for frequency multiplication verification of high-energy harmonics; Step 3: Extract 1 to N of each high-energy harmonic in sequence max The peak value in the frequency multiplication region is based on at least one set of characteristic parameters of the respective amplitudes and / or frequencies, wherein the peak value in the frequency multiplication region is the high energy harmonic itself; and Step 4: Input the at least one set of characteristic parameters of each high-energy harmonic into a machine learning intelligent algorithm for training and calculation.

2. The intelligent identification method according to claim 1, wherein: The at least one set of characteristic parameters includes at least one of the following three sets of characteristic parameters: each high energy harmonic 1 to N max A set of absolute amplitude data, a set of relative amplitude data, and a set of frequency derivative data for the peak value in the octave region.

3. The intelligent identification method according to claim 2, wherein: The set of relative amplitude data refers to a set of bamboo-grass ratio data Or a set of bamboo grass ratio coefficient data The set of frequency derivative data refers to a set of frequency deviation data 4. The intelligent recognition method according to any one of claims 1 to 3, characterized in that: The amplitude comparison in step 2 includes the following sub-steps: Sub-step A: screening of local peaks; Sub-step B, correcting the local peak value and replacing the original value with the corrected value; and Sub-step C: bamboo grass filtering.

5. The identification method according to claim 4, wherein: The bamboo grass filter uses a triangle model or a normal distribution model to allocate weight coefficients of control samples in adjacent windows.

6. The intelligent recognition method according to any one of claims 1 to 3, characterized in that: The upper limit multiple N of the frequency multiplication max The value range is between 3 and 10.

7. The intelligent identification method according to claim 6, characterized in that: The upper limit multiple N of the frequency multiplication max The value range is between 4 and 7.

8. The intelligent identification method as claimed in claim 7, characterized in that: The upper limit multiple N of the frequency multiplication max The value range is between 5 and 6.

9. The intelligent recognition method according to any one of claims 1 to 3, characterized in that: The machine learning intelligent algorithm is a fully connected neural network, including a number of input units and a hierarchical structure distribution that is adapted to the number of the at least one set of characteristic parameters.

10. The intelligent identification method according to any one of claims 1 to 3, characterized in that: The method further comprises step 5: checking whether the frequency (f1) of the high-energy harmonic is consistent with the fault characteristic frequency of the rotating machine.

Images